A Support Vector Based Approach for Classification Beyond the Learned Label Space in Data Streams

Most of the supervised classification algorithms are proposed to classify newly seen instances based on their learned label space. However, in the case of data streams, concept-evolution is inevitable. In this paper we propose a support vector based approach for classification beyond the learned label space in data streams with regard to other challenges in data streams like concept-drift and infinite-length. We maintain the boundaries of observed classes through the stream by utilizing a support vector based method (SVDD). Newly arrived instances located outside these boundaries will be analyzed by constructing neighborhood graph to detect the emergence of a class beyond the learned label space (novel class). Our method is more accurate to model intricate-shape class boundaries than existing method since it utilizes support vector data description method. Dynamically maintaining boundaries by shrinking, enlarging and merging spheres in the kernel space, helps our method to adapt both dramatic and gradual changes of underlying distribution of data, and also be more memory efficient than the existing methods. Conducted experiments on both real and synthetic benchmark data sets show the superiority of the proposed method over the state-of-the-art methods in this area.